mpid_datatype.tex

刚才是说明 现在是安装程序在 LINUX环境下进行编程的MPICH安装文件

\startmanpage
\mantitle{MPID{\tt \char`\_}Datatype}{tex}{10/7/2002}
\manname{MPID{\tt \char`\_}Datatype}
--- Description of the MPID Datatype structure 
\subhead{Synopsis}
\startvb\begin{verbatim}
typedef struct { 
    int           handle;        /* value of MPI_Datatype for this structure */
    volatile int  ref_count;
    int           is_contig;     /* True if data is contiguous (even with 
                                 a (count,datatype) pair) */
    MPID_Dataloop loopinfo;      /* Describes the arguments that the
                                    user provided for creating the datatype;
                                 these are used to implement the
                                 MPI-2 MPI_Type_get_contents functions */
    int           size;
    MPI_Aint      extent;        /* MPI-2 allows a type to created by
                                 resizing (the extent of) an existing 
                                 type.  Note that the extent of the
                                 datatype may be different from the
                                 extent information in loopinfo */

    /* The remaining fields are required but less frequently used, and
       are placed after the more commonly used fields */
    int           has_sticky_ub;   /* The MPI_UB and MPI_LB are sticky */
    int           has_sticky_lb;
    int           is_permanent;  /* e.g., MPI_DOUBLE*/
    int           is_committed;  /* See MPID_Datatype_commit */

    int           loopinfo_depth; /* Depth of dataloop stack needed
                                  to process this datatype.  This 
                                  information is used to ensure that
                                  no datatype is constructed that
                                  cannot be processed (see MPID_Segment) */

    MPI_Aint      ub, lb,        /* MPI-1 upper and lower bounds */
                  true_ub, true_lb; /* MPI-2 true upper and lower bounds */
    MPID_List     attributes;    /* MPI-2 adds datatype attributes */

    int32_t       cache_id;      /* These are used to track which processes */
    MPID_Lpidmask mask;          /* have cached values of this datatype */

    char          name[MPI_MAX_OBJECT_NAME];  /* Required for MPI-2 */

  /* other, device-specific information */
} MPID_Datatype;

\end{verbatim}
\endvb

\subhead{Notes}
The {\tt ref{\tt \char`\_}count} is needed for nonblocking operations such as
\begin{verbatim}
   MPI_Type_struct( ... , &newtype );
   MPI_Irecv( buf, 1000, newtype, ..., &request );
   MPI_Type_free( &newtype );
   ...
   MPI_Wait( &request, &status );
\end{verbatim}

\par
\subhead{Module}
Datatype-DS
\par
\subhead{Notes}
\par
\subhead{Alternatives}
The following alternatives for the layout of this structure were considered.
Most were not chosen because any benefit in performance or memory
efficiency was outweighed by the added complexity of the implementation.
\par
A number of fields contain only boolean inforation ({\tt is{\tt \char`\_}contig},
{\tt has{\tt \char`\_}sticky{\tt \char`\_}ub}, {\tt has{\tt \char`\_}sticky{\tt \char`\_}lb}, {\tt is{\tt \char`\_}permanent}, {\tt is{\tt \char`\_}committed}).  These
could be combined and stored in a single bit vector.
\par
{\tt MPI{\tt \char`\_}Type{\tt \char`\_}dup} could be implemented with a shallow copy, where most of the
data fields, particularly the {\tt opt{\tt \char`\_}dataloop} field, would not be copied into
the new object created by {\tt MPI{\tt \char`\_}Type{\tt \char`\_}dup}; instead, the new object could
point to the data fields in the old object.  However, this requires
more code to make sure that fields are found in the correct objects and that
deleting the old object doesn{\tt t invalidate the dup}ed datatype.
\par
A related optimization would point to the {\tt opt{\tt \char`\_}dataloop} and {\tt dataloop
}fields in other datatypes.  This has the same problems as the shallow
copy implementation.
\par
In addition to the separate {\tt dataloop} and {\tt opt{\tt \char`\_}dataloop} fields, we could
in addition have a separate {\tt hetero{\tt \char`\_}dataloop} optimized for heterogeneous
communication for systems with different data representations.
\par
Earlier versions of the ADI used a single API to change the {\tt ref{\tt \char`\_}count},
with each MPI object type having a separate routine.  Since reference
count changes are always up or down one, and since all MPI objects
are defined to have the {\tt ref{\tt \char`\_}count} field in the same place, the current
ADI3 API uses two routines, {\tt MPIU{\tt \char`\_}Object{\tt \char`\_}add{\tt \char`\_}ref} and
{\tt MPIU{\tt \char`\_}Object{\tt \char`\_}release{\tt \char`\_}ref}, to increment and decrement the reference count.
\par
\endmanpage